We rely on Claude Code (planning) and Codex (review/iteration) . You are the supervisor - you write prompts, drive structure, check outputs, run code, read logs, and iterate.
Core Planning Ritual Inside Claude Code, establish your project foundation with these driver documents:
Create claude.md
Project summary, constraints, target users, KPIs, and explicit tech stack
Create plan.md
Phased implementation plan with tiny milestones and explicit tasks
Create agents.md
Define tools, repos, APIs, and roles for each agent
Use Magic Keys
Apply depth when needed: ultrathink, think deeply, think
Driver Documents claude.md Template # Project: [ Name ] ## Summary [2-3 sentences describing the project] ## Tech Stack (BE EXPLICIT) - Frontend: Next.js 14 + shadcn/ui + Tailwind - Backend: Supabase + Edge Functions - AI: LangChain + LangGraph - Deployment: Vercel ## Constraints - Use battle-tested OSS libraries wherever possible - Avoid custom logic unless necessary - Plan phases so each can be implemented & tested independently ## Target Users [Who will use this and why] ## Success Metrics - [KPI 1] - [KPI 2] plan.md Structure # Implementation Plan ## Phase 0: Foundation (Day 1) - [ ] Scaffold project with Next.js - [ ] Configure environments (.env.local) - [ ] Set up CI/CD pipeline - [ ] Initialize Sentry ## Phase 1: Core Features (Days 2-3) - [ ] Auth flow with Supabase - [ ] Database schema and migrations - [ ] Seed data - [ ] Basic UI shell ## Phase 2: Business Logic (Days 4-5) - [ ] [Feature specific tasks] ## Phase 3: Polish (Day 6) - [ ] E2E tests - [ ] Performance optimization - [ ] Documentation agents.md Roles # Agent Responsibilities ## Claude Code (CC) - Planning & code generation - Architecture decisions - Complex implementations ## Codex - Code review - Refactoring suggestions - Test generation - Security audits ## Runner - Execute commands - Capture logs - Test execution ## Docs Agent - Fetch library examples - API documentation - Best practices Prompt Power-Ups Architecture
Data Modeling
Testing
Refactoring
ultrathink this architecture for edge cases, rate limits, and fallbacks From Plan to Code
Repository Setup
# Create or clone repo git clone [repo-url] cd [repo-name] pnpm i
Branch Strategy
# Feature branches git checkout -b feat/user-auth # Fix branches git checkout -b fix/login-error
Development Loop
# Start dev server pnpm dev # Run tests continuously pnpm test --watch
Tight Iteration
Feed exact errors/logs to CC/Codex: Getting this TypeScript error: [paste exact error] Here's the current code: [paste relevant code] Effective Prompting Good Prompts
Using Next.js 14 app router and Supabase, create a protected dashboard route that: 1. Checks authentication in middleware 2. Fetches user profile with RLS 3. Displays data in shadcn DataTable Include error boundaries and loading states.
This Supabase query returns null: [paste query] RLS policies: [paste policies] Auth context: [paste session info] What's wrong and how do I fix it?
Generate Playwright e2e tests for: - Login flow with valid credentials - Login with invalid credentials - Password reset flow - Session timeout handling Use our existing test helpers from tests/utils.ts Prompt Templates Architecture
Debugging
Review Request
I need to architect a [feature] that: - Requirement 1 - Requirement 2 Current stack: - [List your stack] Constraints: - [Any limitations] Provide: 1. Component structure 2. Data flow 3. Error handling strategy
Working with Codex After Claude Code generates, use Codex for review:
Share Diff
git diff > changes.diff # Paste diff to Codex
Request Review
Review this diff for: - Security issues - Performance problems - Missing error handling - Test coverage gaps
Iterate
Apply suggestions and re-run review until clean
Common Workflows New Feature
Bug Fix
Refactor
Update plan.md with feature requirements
Ask CC to generate implementation plan
Generate code incrementally
Test each piece locally
Codex review → refactor
Write tests
Commit with conventional message
Best Practices Always specify your tech stack explicitly - don’t let AI guess what libraries you’re using
Use battle-tested libraries - explicitly tell AI to prefer proven OSS over custom implementations
Small, atomic prompts work better than large, complex ones. Break down big tasks.
Troubleshooting
AI generates wrong library usage
Always specify exact versions and import styles: Using Next.js 14 with app router (not pages), Supabase JS client v2 (not v1), shadcn/ui components from @/components/ui
Generated code doesn't match style
Provide examples from your codebase: Follow this pattern from our codebase: [paste existing code example]
Complex logic gets confused
Break into smaller steps: Step 1: Just create the database schema Step 2: Now add the RLS policies Step 3: Create the API route 
